The DLG collector as used by Caml Light does what you describe. Updating an object in the global heap to point to a thread-local heap can cause unbounded work

The DLG collector as used by Caml Light does what you describe. Updating an object in the global heap to point to a thread-local heap can cause unbounded work to be done before that update as the entire pointed-to structure has to be copied into the global heap first. GC designers really hate write barriers (sections of code that are performed before a pointer is updated) that perform unbounded work.